The Divergent Autoencoder (DIVA) Account of Human Category Learning

Page created by Juan Avila
 
CONTINUE READING
The Divergent Autoencoder (DIVA) Account of Human Category Learning
                                          Kenneth J. Kurtz (kkurtz@binghamton.edu)
                                              Department of Psychology, PO Box 6000
                                        Binghamton University (State University of New York)
                                                   Binghamton, NY 13902 USA

                            Abstract                                       to n-way classification tasks or to cases where an
                                                                           A/B/neither classification response is required.
The DIVA network model is introduced based on the novel                       The innovation unique to DIVA is a method for
computational principle of divergent autoencoding. DIVA                    converting any supervised learning problem into a form
produces excellent fits to classic data sets from Shepard,
                                                                           addressable by autoassociative learning. Traditionally, an
Hovland & Jenkins (1961) and Medin & Schafffer (1978).
DIVA is also resistant to catastrophic interference. Such results
                                                                           autoassociative system is only capable of categorization to
have not previously been demonstrated by a model that is not               the extent that it picks out the statistical structure of a
committed to both localist coding of exemplars (or exceptions)             training set in a manner like clustering. This process
and the use of an explicit selective attention mechanism.                  suggests category formation in the sense that if a training
                                                                           environment is naturally organized in terms of sets of self-
                         Introduction                                      similar cases, the autoassociative learning system will
  The problem of supervised classification learning is of                  extract that structure. Similar inputs are similarly
fundamental importance in both cognitive psychology and                    represented and subsequent generalization behavior reflects
machine learning. Models of many kinds have been put                       these attractors. However, such a system has no capacity to
forward offering powerful solutions. This paper presents a                 acquire a classification scheme based on supervision that
novel approach to supervised learning that shows                           crosscuts the correlational structure of the training set.
considerable promise as an account of human category                          The computational principle of divergent autoencoding
learning and as a technology for applied problems. The                     offers an elegant solution to this problem using an
DIVergent Autoencoder (DIVA) network model takes as a                      autoassociative learning channel for each output class in an
starting point the back-propagation learning algorithm                     n-way classification problem. For a standard A/B
(Rumelhart, Hinton, & Williams, 1986) and the                              classification learning task, one output channel is designated
reconstructive autoencoder architecture (McClelland &                      for reconstructing patterns labeled A by the teaching signal
Rumelhart, 1986). Autoassociative systems are powerful                     and the other is assigned to patterns labeled B. No output
learning devices that have been shown to implement                         units are explicitly assigned to code for the categories
principle component analysis and avoid local minima (Baldi                 themselves. The correct classification choice is used to
& Hornik, 1989); to be extensible to non-linear function                   select the channel on which to apply the targets (which are
approximation (Japkowicz, Hanson, & Gluck, 2000); and to                   the same as the input). The architecture consists of an input
perform compression (e.g., DeMers & Cottrell, 1993).                       layer, a shared hidden layer, and a set of autoassociative
DIVA also draws on a design principle of multi-task                        output banks. The pattern of connectivity is full and
learning mediated through a common hidden layer that been                  feedforward; all weight update is by back-propagation.
articulated in the ORACL model of concept formation
(Kurtz, 1997; Kurtz & Smith, in preparation) as well as in                            Autoencoder A              Autoencoder B
the literature on neural computation (Caruana, 1995; Intrator
& Edelman, 1997; Gluck & Myers, 1993).
  Japkowicz (2001) developed an approach for applying
unsupervised learning to binary classification that is close in
spirit to the present proposal. An autoencoder is trained only
on the positive instances of a category. Subsequently, inputs
can be tested for membership in the category by evaluating                                                            Shared
the reconstructive success of the autoencoder. A new                                                                  Hidden Units
example that is consistent with the set of learned category
examples will show minimal error while an example that is
inconsistent will show a higher level of output error
suggesting an inability to construe the input as a category                                                                Input
member. Japkowicz (2001) demonstrated good results on                                                                      Features
binary classification problems by training a model to
recognize examples of one class. Successful reconstructions
are classified as members and rejections are assumed to                             Figure 1: Architecture of the DIVA network.
belong to the other category. The approach is not extensible

                                                                    1214
The recoding of input information at the hidden layer of          roughly equivalent performance of FR, Type III, and Type
DIVA is shared by the set of channels, each of which is              V; followed by Type VI (though see Kurtz, under review).
dedicated to learning to reconstruct the members of one                 The relative ease of learning the six types was tested
class. This is different from forming compressed                     across six random initializations of a (3-2-3x2) DIVA
representations in traditional autoencoding and also differs         network (note: this refers to a DIVA network with three
from learning a recoding to achieve a linearly separable             input units, two hidden units, and two autoassociative output
boundary between classes in a standard multi-layer                   channels with three units each). The number of epochs to
perceptron architecture. DIVA will tend to produce different         criterion was determined based on total sum-squared error
internal representations of an item depending upon the other         across the eight training patterns. Error was recorded only
same-category members included in the training set and also          on the target-active (correct) channel. Two stopping points
depending upon the contrasting categories being learned at           (SSE = .2; SSE = .1) were applied in accord with the strict
the same time.                                                       criteria used in the behavioral study.
   The DIVA network is tested by presenting an input which              For the data reported in Table 1, learning rate of 0.25 and
is processed along each channel in parallel. A classification        initial weight range of zero +/- 0.5 were used. However,
response can be based on the amount of reconstructive error          qualitative performance was found to be consistent across
along a particular channel (i.e., testing the hypothesis that        variations in learning rate and the range of initial weight
the example is a member of a particular category) as in              randomization. The only critical parameter is the number of
Japkowicz (2001). In standard n-way classification tasks,            hidden units. A simple systematic basis is used to determine
the response is determined by selecting the class                    the number of hidden units for a task. The smallest number
corresponding to the channel with the best reconstruction,           of hidden units that can successfully reach asymptotic
i.e., the lowest sum-squared error. A version of Luce’s              minimization of error across the manipulated learning
(1963) choice rule is used to generate response probabilities        conditions is the number that are used. This approach is in
for each choice K based on the inverse of the sum squared            sharp contrast to the usual technique of exhaustive search
error at the output layer of the N channels. This is an              through parameter space to find the best fit for each
extension of the common application of the choice rule to            phenomenon of interest. In this case, two hidden units were
response generation based on output unit activations (e.g.,          required to consistently reduce error on the six SHJ types.
Kruschke, 1992):
                                                                          Table 1: Relative ease of category learning by DIVA
                               N

       Pr (K) = (1/SSE(K)) /   Σ (1/SSE(k)),            (1)
                                                                           SHJ Type
                                                                                            Mean number
                                                                                            of Epochs to
                                                                                                              Mean number
                                                                                                              of Epochs to
                               k=1
                                                                                            criterion (0.2)   criterion (0.1)
                                                                           I                      566               840
  The logic of this paper is to demonstrate the power and
                                                                           II                     847              1295
promise of the DIVA network for cognitive simulation. To
address the topic of human category learning, the primary                  III                   1195              1953
goal is to evaluate the model on the two most widely studied               IV                    1232              2087
datasets in the literature: the Shepard, Hovland, & Jenkins’               V                     1144              1750
(1961) dataset on ease of learning and the 5-4 learning                    VI                    5719              9416
problem introduced by Medin & Schaffer (1978). In
addition to fitting benchmark data, a number of appreciable             As can be seen in Table 1, the data are well fit (Type 1 <
properties of DIVA in comparison with competing models               Type II < Types III, IV ,V < Type VI). Consistent findings
will be outlined.                                                    were observed across the time course of training as was
                                                                     found in the SHJ replication by Nosofsky, Gluck, Palmeri,
  Experiment 1. The Relative Ease of Learning                        McKinley, & Glauthier (1994). By way of comparison, a
         Across Category Structures                                  standard feedforward (4-2-1) back-propagation network was
                                                                     tested under matching conditions. As also reported by
Shepard, Hovland, & Jenkins (1961) produced a
                                                                     Kruschke (1992), the network was far too quick to learn FR
groundbreaking analysis of the rate of acquisition of the six
                                                                     (comparable speed to UNI) and too slow to learn XOR.
general types of category structures that are possible within
                                                                     With two hidden units, some initializations became stuck in
a training set of binary-valued, overtly analyzable, three-
                                                                     local minima (especially on Type V) and the system showed
dimensional stimuli. The most interpretable of these                 no progress on Type VI (a version of parity problem)
structures are: Type I, a unidimensional rule (UNI); Type II,        without more hidden units.
the exclusive-or problem plus an irrelevant dimension                   Another way to test the performance of DIVA is to
(XOR); and Type IV, a family resemblance structure (FR).             compute the classification response to each pattern using the
The results that have generated considerable challenges to           choice rule over sum-squared error as outlined above. A
model-builders is a qualitative ordering of the relative ease
                                                                     single simulation was conducted in this fashion using the
of learning: UNI fastest; followed by XOR; followed by

                                                              1215
identical set of initial weights for each of the SHJ types. The       explicitly represented the critical correlation between the
learning results after 500 training epochs appear in Table 2.         diagnostic features (a standard back-propagation network
                                                                      would search for a recoding of the input specifically
       Table 2: Classification accuracy for a DIVA network.           targeted to allow for linearly separable classification
                                                                      between the hidden layer to the output.) F1 and F3 were
         SHJ Problem    Category          Classification              always correct on the ‘incorrect’ category channel, and the
         Type           Structure         Accuracy                    output there for F2 was always exactly opposite to the input
         I              UNI                   .97                     activation.
         II             XOR                   .94                       On the FR problem, a representative DIVA network
         III                                  .84                     reached the following solution. H1 received an excitatory
         IV             FR                    .83                     signal from F3 and an inhibitory signal from F2. H2 was
         V                                    .93                     sensitive to all three input features with a strong inhibitory
                                                                      signal from F1 and lesser excitation from F2 and F3
         VI                                   .56
                                                                      yielding the recodings shown in Table 4.
  The fit is excellent except for the overly good                         Table 4: Recodings formed by a DIVA network on FR.
performance of Type V. There is a degree of variation
across initializations of DIVA networks and in the case                Input        Hidden1        Hidden2          Target
presented above, Type V showed performance at the upper                             Activation     Activation       Category
end of its usual range. Such variation results primarily from
                                                                       101          1              0                1
the degree of consistency between the initial random
                                                                       001          1              0.9              0
configuration of weights and the form of the solution that is
                                                                       000          0.4            0.6              0
required. When lower learning rates and smaller initial
                                                                       011          0.5            1                1
weight variation are selected, the degree of variation lessens
                                                                       111          0.5            0.4              1
considerably.
                                                                       100          0.5            0                0
  In order to make clear how the learning occurs, DIVA
                                                                       110          0              0                1
solutions to the most interesting of the SHJ problem types
                                                                       010          0              1                0
(UNI, XOR, FR) are described as follows. A representative
DIVA network solved the UNI problem (on F1) by
                                                                      The network assigned each input item to a unique location
assigning one hidden unit to code for the presence of F1 and
                                                                      in the two-dimensional representational space of the hidden
F2 respectively. Each hidden unit strongly activated the F1
                                                                      layer. The two channels showed equivalent connectivity
output units: via excitation on one channel and inhibition on
                                                                      projecting from the hidden layer and used strong bias
the other. Each hidden unit also activated the appropriate
                                                                      weights to differentiate their performance. It is interesting to
non-diagnostic feature on each channel. F2 and F3 were
                                                                      note that this solution parallels the behavior of an ordinary
always correct on the ‘incorrect’ category channel, while the
                                                                      autoencoder operating on this training set. Once again,
output there for F1 was always exactly opposite to the input
                                                                      while operating entirely on the basis of the back-
activation.
                                                                      propagation algorithm, the hidden units do not act to
  To solve XOR (on F1 and F2), a representative DIVA
                                                                      transform the input for linearly separable classification. The
network largely ignored F2, but used signals from F1 and
                                                                      ‘incorrect’ channel attempts to interpret each input as a
F3 to generate hidden layer recodings as shown in Table 3.
                                                                      member of its category and therefore produces markedly
    Table 3: Recodings formed by DIVA network on XOR.                 increased or reducing activation on one or more of the
                                                                      features.
                                                                         The XOR problem holds a high place in the contemporary
 Input         Hidden1         Hidden2            Target
                                                                      study of both human and machine learning. For decades, the
               Activation      Activation         Category
                                                                      connectionist tradition was halted by the lack of an
 101           0.2             0                  0
                                                                      algorithm to handle cases of hard learning, i.e., non-linearly
 001           0               0.8                1
                                                                      separable functions. Rumelhart, Hinton, & Williams’ (1986)
 000           0.8             1                  0
                                                                      paper on back-propagation of errors was a breakthrough that
 011           0               0.9                0
                                                                      elicited tremendous productivity. The XOR problem
 111           0.2             0                  1
                                                                      remains a benchmark for evaluation of learning systems. A
 100           1               0                  0
                                                                      standard (hetero-associative) back-propagation network
 110           1               0                  1
                                                                      reaches asymptote on Type II learning (the XOR problem
 010           0.8             1                  0
                                                                      with an added irrelevant dimension) after approximately
                                                                      3000 epochs of training. The DIVA network reached
The DIVA network used four areas of the activation space              asymptote on average in 847 epochs. This nearly fourfold
on H1 to code for the pairwise combinations of the                    increase in speed of learning suggests that DIVA can
diagnostic F1 and the non-diagnostic F3, while H2 primarily           perform non-linear function approximation with
coded for F1. It is interesting to note that neither hidden unit

                                                               1216
considerable ease. The SHJ Type VI is the parity problem                from the prototype. Once again, the advantage goes to the
with three dimensions. The standard back-propagation                    exemplar view.
network did not make any headway with two hidden units,                    A (4-2-4x2) DIVA network was applied to the 5-4
but the DIVA network coasted smoothly down the error                    problem using a learning rate of 0.1 and initial weights
gradient. These findings suggest the power of the DIVA                  randomized in a range of zero +/- .05. The model was
network as a general learning device.                                   allowed to run for 1000 epochs. Performance on each
   In sum, the Shepard, Hovland, & Jenkins (1961) dataset is            training instance and the transfer items was determined by
something of a litmus test for models of classification                 applying the choice rule to the sum-squared error along each
learning. Despite some question about the generality of the             channel. In terms of quantitative fit, a correlation of .96 was
finding (see Kurtz, under review), it a seminal result in the           found between the probabilistic responses of the DIVA
literature. The design features of those models which have              network and a summarization of thirty different behavioral
successfully fit this data have come to represent the state of          tests of the 5-4 problem published by Smith & Minda
the art in the field. Localist encoding and selective attention         (2000). The DIVA network produced a probability of A,
are core components of the three successful models:                     Pr(A) = .96 for Stimulus A2 and Pr(A) = .85 for Stimulus
ALCOVE (Kruschke, 1992), SUSTAIN (Love, Medin, &                        A1; thereby fitting the critical qualitative result that was
Gureckis, 2004) and RULEX (Nosofsky, et al., 1994).                     previously captured only by pure exemplar models and
These models all depend upon multiple free parameters (not              RULEX (Nosofsky, Palmeri, & McKinley, 1994). In
including learning rate) that are selected according to the             addition, the transfer item T3 which is the prototype of
same data that is to be fit. RULEX uses three best-fitting              Category A produced Pr(A) = .86 which was the strongest
parameters in addition to best-fitting attentional weights.             response to any transfer item, but was a lesser response than
ALCOVE and SUSTAIN each use three best-fitting                          that shown for the training items A2 and A3. DIVA offers
parameters. DIVA offers a successful fit with a single                  the first successful fit to these results by a model that does
parameter which is set a priori, rather than post-hoc, and              not implement the theoretical framework of localist
offers a strong challenge to the widespread view that                   encoding and selective attention.
selective attention and localist representation are the correct
explanatory constructs.                                                        Experiment 3. Avoiding Catastrophic
                                                                                         Interference
           Experiment 2. Learning the 5-4                               Among some researchers, the phenomenon of catastrophic
             Categorization Problem                                     interference has been considered a fatal flaw for back-
The case for the superiority of exemplar models has rested              propagation as an account of human learning and memory
in no small part on extensive behavioral and computational              (e.g., McCloskey & Cohen, 1989). In point of fact, a
tests of the 5-4 problem introduced by Medin & Schaffer                 number of intriguing solutions and more nuanced treatments
(1978). A challenge has been raised recently (e.g., Smith &             (McClelland, McNaughton, & O’Reilly, 1995; Mirman &
Minda, 2000) based on successful fits by a ‘souped-up’                  Spivey, 2001) have appeared. Nonetheless, a minimal
version of a prototype model and questioning of the                     solution (one that does not graft an additional component,
satisfactory nature of the exemplar account presented by                integrate additional mechanisms, or make modifications to
Nosofsky, Kruschke, & McKinley (1992).                                  the training set, etc.) has not been found. Is it possible to
   The 5-4 category problem consists of nine training items             preserve the computational power and psychological
with four binary-valued features plus a set of transfer items.          validity of learning distributed internal representations via
The design feature of the problem is that it is linearly                back-propagation without catastrophic interference?
separable (and therefore fair game for testing prototype                  The definitive demonstration of catastrophic interference
models), but includes three very weak category members                  for neural network models trained by back propagation is
(for which only two out of the four features are consistent             Ratcliffe’s (1990) simulation result using the 4-4 encoder
with the underlying prototype). Category B consists only of             problem. The problem involves two learning phases.
its prototype, one strong example, and two weak examples.               Training is performed to a certain level on the Phase I
   Model testing has focused not only on overall quantitative           examples and then the training set is swapped. Phase II
fit, but also to two qualitative aspects of the data. The first is      consists of training on only the second training set. The
that in non-elaborated experimental versions of the task,               observed phenomenon is that the network performs well on
learners are more accurate on Stimulus A2 (which has two                the first training set at the end of Phase I, but the process of
features in common with the A prototype) than they are on               learning in Phase II “catastrophically” disrupts performance
Stimulus A1 (which has three prototypical features). The                on Phase I examples. Phase I consists of three four-
prototype model predicts the opposite, while exemplar                   dimensional patterns to be autoassociatively reconstructed
models capture the result (Nosofsky, et al., 1992). In                  through an intermediate hidden layer. The patterns are:
addition, behavioral results typically show that a transfer             1000, 0100, and 0010. Phase II consists of a single pattern:
test on the Category A prototype produces highly accurate               0001.
responding, though not more so than the observed                          Using DIVA, it is straightforward to assign a separate
performance on training items that are somewhat distant                 output channel to each sequential phase of learning. The

                                                                 1217
divergent autoencoding principle is applied in this case to            are hardly affected by Phase II training, and the weights
separate phases of learning rather than to separate                    from the hidden layer to the P1 channel are affected not at
classification labels (as above). The same input and hidden            all. However, this is not at all equivalent to using entirely
units are used, however separate bank of outputs are used              different networks for the two phases of learning. The same
for each phase. Both channels are present in the architecture          input units, hidden units, and connecting weights are used.
at all times, but targets only are applied to adjust the weights       The two learning phases are equivalent for DIVA to
along the active channel. The critical assumption is that the          learning a two-way classification problem with massed
shift between phases of learning must somehow be                       practice. One can interpret the DIVA solution to the
demarcated and psychologically encoded. The task context               problem of catastrophic interference as the establishment of
must make clear that “now you are to learn something else.”            a contextually-driven classification of inputs as members of
In point of fact, traditional paradigms for studying                   either Phase 1 or Phase 2. With this one very plausible
interference usually make a very clear distinction between             assumption, divergent autoencoding preserves the back-
List 1 and List 2. An intriguing prediction is that an                 propagation machinery for error-driven learning without the
unannounced or non-obvious shift from Phase I to Phase II              catastrophic interference.
ought to elicit CI unless the switch is made manifest. As a
final point of emphasis, no known model has been able to                                  General Discussion
exploit the phase variable to prevent CI by devoting input or            Given the demonstrated promise of DIVA, a number of
output units to code for the phase of each presented pattern.          further explorations are underway. DIVA shows a tendency
   A (4-3-4x2) DIVA network was tested with three hidden               to shift during learning from more general to more specific
units and a learning rate of 0.2 in accord with Ratcliffe              category representations (e.g., Smith & Minda, 1998).
(1990) and Kruschke (1992). Weights were randomly                      DIVA is naturally extensible to the recently vigorous
initialized in a tighter range around zero. The network                investigation of category learning beyond traditional
required 550 epochs to reach the 70% training criterion for            classification, i.e., inference learning, category use,
Phase 1 learning used by previous investigators. As                    unsupervised learning, and cross-classification. Since
explained above, Phase I training applied targets only on the          autoassociative processing naturally generates a feature-
P1 channel. The same amount of training was conducted for              based representation as its output, applications to
Phase II on just the 0001 pattern using only the targets on            recognition memory, memory distortions, and feature
the P2 channel.                                                        prediction are forthcoming.
                                                                         An intriguing aspect of the DIVA architecture is that it
       Table 5: Output Activations of DIVA network on                  offers a straightforward mechanism for producing a
                 Sequential Learning Task.                             convolved representation of any input in terms of any
                                                                       category known to the network. Imagine that a pattern
   Input               Channel for P1        Channel for P2
                                                                       representing a cat is presented to a DIVA network trained
 After Phase 1
                                                                       on various animal concepts. Regardless of which animal is
   1000                .74 .19 .17 .04
   0100                .18 .68 .23 .03                                 the actual classification response, every channel produces an
   0010                .16 .24 .70 .04                                 interpretation or construal of the input in terms of its
   0001                                       .49 .49 .49 .50          category. The psychological nature of such construals is of
 After Phase 2                                                         great interest. For example, the similarity of concepts A and
   1000                .73 .21 .15 .03                                 B can be computed as the degree of reconstructive success a
   0100                .16 .66 .24 .03                                 DIVA network achieves in processing a prototypical
   0010                .17 .22 .68 .03                                 example of A along a channel trained on concept B.
   0001                                       .04 .04 .04 .96          Typicality or graded structure of category members can be
                                                                       understood as the degree of reconstructive success in
   As shown in Table 5, catastrophic forgetting was fully              processing a member of category A through the channel for
avoided. Similar performance was observed across                       that category. Argument strength for category-based
differently initialized runs and variations in learning rate           induction can be understood as the degree of reconstructive
and initial weight range. Two follow-up tests were                     success in processing a representation of the conclusion
conducted. The DIVA network was tested using negative                  category along the channel(s) of premise categories. The
valued (-1) input activations rather than zero-valued ‘off’            internal representation generated by inputting a
units. This also yielded successful results. In addition, an           representation or representative example of one concept to
alternate version of Phase II learning was conducted using             the channel of another concept is likely to produce a
the pattern . This extends the problem beyond               conceptual combination or metaphoric interpretation. If a
the case in which positive activation of the features is               parsimonious means can be found to represent structural
segregated between the two training phases. Once again,                information in a form submittable to a neural network, the
performance on Phase I examples remained intact.                       potential deepens.
   The success of the DIVA network can be explained very
simply. The weights from Features 1-3 to the hidden layer

                                                                1218
In sum, DIVA provides an uncompromisingly good fit to             mathematical psychology (pp. 103-189). New York:
the two most influential data sets on human category                Wiley.
learning and does so with the following characteristics:          McClelland, J.L., McNaughton, B.L. & O'Reilly, R.C.
  1. Distributed representation rather than localist nodes          (1995). Why There are Complementary Learning Systems
       for individual instances                                     in the Hippocampus and Neocortex: Insights from the
  2. No selective attention mechanism                               Successes and Failures of Connectionist Models of
  3. No performance-optimized free parameters                       Learning and Memory. Psychological Review, 102, 419-
Therefore, the success of this model calls into question            457.
widely held theoretical assumptions. The DIVA network             McClelland, J.L. & Rumelhart, D.E. (1986). A Distributed
offers the brain-style computational power of back-                 Model of Memory. In Rumelhart, D. E., McClelland, J.L.
                                                                    (Eds.), Parallel distributed processing: Explorations in
propagation and overcomes its shortcomings in simulating
                                                                    the microstructure of cognition: Vol 1I. Applications
human learning. The computational design principle of
                                                                    (pp.170-215). Cambridge MA: MIT Press.
divergent autoencoding deserves consideration as an               McCloskey, M., & Cohen, N. J. (1989). Catastrophic
explanatory construct underlying broad aspects of cognition.        interference in connectionist networks: The sequential
                                                                    learning problem. In G. H. Bower (Ed.), The psychology
                   Acknowledgments                                  of learning and motivation (Vol 24, pp. 109-165). New
With thanks to David E. Rumelhart. This project was                 York: Academic Press.
partially supported by NIH award 1R03MH68412-1.                   Medin, D. L., & Schaffer, M. M. (1978). Context theory of
                                                                    classification learning. Psychological Review, 85, 207-
                       References                                   238.
                                                                  Mirman, D. and Spivey, M. (2001) Retroactive interference
Baldi, P. and Hornik, K. (1989) Neural networks and                 in neural networks and in humans: the effect of pattern-
  principal components analysis: Learning from examples             based learning. Connection Science, 13(3), 257-275.
  without local minima. Neural Networks, 2, 53-58.                Nosofsky, R.M., Gluck, M., Palmeri, T.J., McKinley, S.C.,
Caruana, R. (1995). Learning many related tasks at the same         & Glauthier, P. (1994). Comparing models of rule-based
  time with backpropagation, Advances in Neural                     classification learning: A replication and extension of
  Information Processing Systems Vol. 7, pp. 657-664,               Shepard, Hovland, and Jenkins (1961). Memory &
  Morgan Kaufmann, San Mateo, CA, 1995.                             Cognition, 22, 352-369.
DeMers, D. & Cottrell, G. (1993). Nonlinear dimensionality        Nosofsky, R., Kruschke, J., & McKinley, S. (1992).
  reduction. Advances in Neural Information Processing              Combining exemplar-based category representations and
  Systems Vol. 5, pp. 580-587, San Mateo, CA: Morgan                connectionist learning rules. Journal of Experimental
  Kaufmann.                                                         Psychology: Learning, Memory, and Cognition, 18, 211-
Gluck, M. A. & Myers, C. E. (1993). Hippocampal                     233.
  mediation of stimulus representation: A computational           Nosofsky, R. M., Palmeri, T. J., & McKinley, S. K. (1994).
  theory. Hippocampus, 3, 491-516.                                  Rule-plus-exception model of classification learning.
Intrator, N. and Edelman, S. (1997). Learning low-                  Psychological Review, 101,55-79.
  dimensional representations via the usage of multiple-          Ratcliff, R. (1990). Connectionist models of recognition
  class labels. Network 8, 259-281.                                 memory: Constraints imposed by learning and forgetting
Japkowicz, N. (2001). Supervised versus unsupervised                functions. Psychological Review, 97, 285-308.
  binary-learning by feedforward neural networks. Machine         Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986).
  Learning, 42, 97-122.                                             Learning internal representations by error propagation. In
Japkowicz, N., Hanson S.J., & Gluck, M.A. (2000).                   D. E. Rumelhart & J. L. McClelland (Eds.), Parallel
  Nonlinear Autoassociation is not Equivalent to PCA.               distributed processing: Explorations in the microstructure
  Neural Computation, 12, 531-545.                                  of cognition: Vol 1. Foundations (pp.318-362).
Kruschke, J. K. (1992). ALCOVE: An exemplar-based                   Cambridge, MA: Bradford Books/MIT Press.
  connectionist model of category learning. Psychological         Shepard, R.N., Hovland, C.L., & Jenkins, H.M. (1961).
  Review, 99, 22-44.                                                Learning and memorization of classifications.
Kurtz, K.J (1997). The influence of category learning on            Psychological Monographs, 75 (13, Whole No. 517).
  similarity. Unpublished doctoral dissertation.                  Smith, J. D., & Minda, J. P. (1998). Prototypes in the mist:
Kurtz, K.J. (under review). Abstraction versus selective            The early epochs of category learning. Journal of
  attention in classification learning.                             Experimental Psychology: Learning, Memory, and
Kurtz, K.J. & Smith, G. (in preparation). The ORACL                 Cognition, 24, 1411-1430.
  model of concept formation and representation.                  Smith, J.D. & Minda, J.P. (2000). Thirty categorization
Love, B.C., Medin, D.L, & Gureckis, T.M (2004).                     results in search of a model. Journal of Experimental
  SUSTAIN: A Network Model of Category Learning.                    Psychology: Learning, Memory, and Cognition, 26, 3-27.
  Psychological Review, 111, 309-332.
Luce, R.D. (1963). Detection and recognition. In R.D. Luce,
  R.R. Bush, & E. Galanter (Eds.), Handbook of

                                                           1219
You can also read